Voxblox
Voxblox 以體素雜湊(voxel hashing)儲存 TSDF,並提出兩項整合策略:同一體素內的點先分組取加權平均再只射線投射一次(grouped raycasting),以及考量深度平方雜訊與表面後方線性衰減的權重函數。接著以波前傳播(raise/lower wavefront)由 TSDF 增量建立歐氏有號距離場(ESDF)供無人機路徑規劃,並可隨時以 marching cubes 輸出網格。
本頁內容
Builds TSDFs with voxel hashing, grouped ray casting and a depth-dependent weighting with behind-surface drop-off, then incrementally derives an ESDF for planning and meshes on demand.
技術屬性
欄位內容為文獻擷取紀錄的原文用語(英文),以原文為據;「未查證」表示本研究尚未讀到該資訊,不代表該方法不具備此能力。
| 感測輸入 | RGB-D、stereo |
|---|---|
| 原文測試平台 | UAV (AscTec Firefly with forward stereo camera and IMU, on-board experiment; EuRoC MAV V1_01 dataset)、simulation (noiseless RGB-D, 320 x 240, Sec. VI-B1) |
| 狀態估計 | 不適用 (poses supplied externally, e.g., Vicon or visual-inertial pose) |
| 資料關聯 | grouped (merged) ray casting of points into TSDF voxels |
| 時間表示 | 不適用 |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | voxel-hashed TSDF layer, incremental ESDF layer, mesh layer |
| 先驗資訊 | external poses |
| 可輸出幾何 | TSDF, ESDF and incremental marching-cubes mesh |
| 計算需求 | single CPU thread; experiments on a quad-core i7 at 2.5 GHz using one thread; on the MAV an on-board Intel i7 at 2.1 GHz ran state estimation, TSDF integration, ESDF updates at 4 Hz with 0.20 m voxels and replanning within a 250 ms budget (Sec. VI, VII) |
使用設備
原文使用的感測器、運算硬體與載具(equipment)。型號保留原文寫法,連結到設備頁中同一型號的歸併名稱;角色依原文用途分為方法輸入、資料集感測器、執行運算平台、參考或真值量測(reference or ground truth)與比較對象設備。
| 類別 | 型號(原文寫法) | 角色 | 資料集 | 原文規格 | 出處 |
|---|---|---|---|---|---|
| 慣性量測單元(IMU) | IMU synced to the stereo camera (model not named) | 方法輸入 | 未標示 | input to the visual-inertial state estimator | (Oleynikova et al., 2017, Sec. VII) |
| 雙目相機 | narrow-baseline grayscale stereo sensor (model not named) | 資料集感測器 | EuRoC MAV (V1_01_easy) | 原文未報告 | (Oleynikova et al., 2017, Sec. VI) |
| 雙目相機 | forward-facing stereo camera synced to an IMU (model not named) | 方法輸入 | 未標示 | stereo matching input to mapping and to the visual-inertial estimator | (Oleynikova et al., 2017, Sec. VII) |
| RGB-D 相機 | Microsoft Kinect (original) | 資料集感測器 | cow dataset | 原文未報告 | (Oleynikova et al., 2017, Sec. VI) |
| 全測站 | Leica TPS MS50 | 參考或真值量測 | cow dataset | described as a laser scanner; 3 scans merged for structure ground truth | (Oleynikova et al., 2017, Sec. VI) |
| 全測站 | Leica TPS MS50 | 參考或真值量測 | EuRoC MAV (V1_01_easy) | scans used as structure ground truth | (Oleynikova et al., 2017, Sec. VI) |
| 載具平台 | AscTec Firefly | 方法輸入 | 未標示 | MAV; all estimation, mapping, planning and control on board | (Oleynikova et al., 2017, Sec. VII, Fig. 1) |
| 運算硬體 | quad-core Intel i7 2.5 GHz | 執行運算平台 | 未標示 | one thread used | (Oleynikova et al., 2017, Sec. VI) |
| 運算硬體 | Intel i7 2.1 GHz (on-board) | 執行運算平台 | 未標示 | on-board MAV computer | (Oleynikova et al., 2017, Sec. VII) |
| 其他 | Vicon motion capture system | 參考或真值量測 | cow dataset | pose source for the cow dataset | (Oleynikova et al., 2017, Sec. VI) |
| 其他 | Vicon fused with IMU | 參考或真值量測 | EuRoC MAV (V1_01_easy) | pose information for EuRoC | (Oleynikova et al., 2017, Sec. VI) |
| 其他 | simulated noiseless RGB-D sensor | 資料集感測器 | synthetic ESDF benchmark | 320 x 240, maximum range 5 m, 50 random poses | (Oleynikova et al., 2017, Sec. VI-B1) |
作者報告的優勢與限制
優勢
- Grouped ray casting is up to 20 times faster than standard ray casting into a TSDF and up to 2 times faster than grouped OctoMap insertion (Sec. IV-B, Fig. 6).
- Proposed quadratic weighting yields lower reconstruction RMS error than constant weighting, with a larger effect at larger voxel sizes (Sec. VI-A, Fig. 5).
- In simulation every TSDF-derived ESDF variant had lower error than an ESDF built from occupancy, and a one-voxel fixed band gave the lowest error (Sec. VI-B1, Fig. 8).
- Incremental ESDF updates were about an order of magnitude faster than batch updates on EuRoC (Sec. VI-B2, Fig. 9).
限制
- Quasi-Euclidean ESDF distances introduce error; authors recommend inflating robot bounding boxes by 8.25% (Sec. V-B).
- Full Euclidean distance reduced ESDF error by 8.23/5.18/4.72% at 0.05/0.10/0.20 m voxels but increased integration time by 201.0/61.3/33.9% (Sec. VI-B).
- nvblox authors state that voxblox is limited in achievable map resolution by CPU update cost (Millane et al. 2024, Sec. II).
- The conclusion recommends a more conservative inflation of 8.5% + 0.3v (v = voxel size) to cover map errors (Sec. VIII); cite the section-specific value.
營建工程相關證據
未在營建場域驗證;TSDF 表面誤差以 Leica TPS MS50 雷射掃描所得結構真值評估(cow 資料集與 EuRoC V1_01),並報告誤差隨體素尺寸與權重函數而變。
原文驗證環境:公開基準、受控實驗、模擬、獨立參考量測
報告的性能數據
以下是原文作者報告的性能數值(author-reported results),不是本研究重新量測的結果。每張圖只並列同一個比較組(comparison group,同一張表、同一組實驗設定)內的方法;不同比較組之間的數值不可直接比較,也不構成排名。
本方法共出現在 14 個比較組,合計 96 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 10 組列在最後,並連到性能比較頁。
Millane et al., 2024 · Table I 本方法 30 筆
表格設定(擷取紀錄原文):Component runtimes averaged over 8 Replica and 5 Redwood sequences at 5 cm voxels; ESDF and mesh computed every 4 frames; voxblox does not separate TSDF and color integration, so its TSDF and Color cells are empty; speed-up column not extracted (Millane et al., 2024, Table I)
ESDF component runtime (ms), computed every 4 frames,Replica · average over sequences
只並列這張表在相同設定下報告的方法;以「本方法:」開頭者為本頁方法。失敗、未執行與未報告以標記呈現,不是 0。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Millane et al., 2024 在此表設定下報告的數值(author-reported results),只能在同一個比較組內對照,不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Millane et al., 2024, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| nvblox(Desktop: Intel i9 CPU, NVIDIA RTX 3090 Ti)原文提出 | 1.9 ms | (Millane et al., 2024, Table I) |
| nvblox(Laptop: Intel i7 CPU, RTX 3000 Mobile GPU)原文提出 | 3.6 ms | (Millane et al., 2024, Table I) |
| nvblox(NVIDIA Jetson AGX Xavier)原文提出 | 8.4 ms | (Millane et al., 2024, Table I) |
| voxblox(Desktop: Intel i9 CPU, NVIDIA RTX 3090 Ti)本方法 | 163.2 ms | (Millane et al., 2024, Table I) |
| voxblox(Laptop: Intel i7 CPU, RTX 3000 Mobile GPU)本方法 | 291.5 ms | (Millane et al., 2024, Table I) |
| voxblox(NVIDIA Jetson AGX Xavier)本方法 | 231.6 ms | (Millane et al., 2024, Table I) |
Millane et al., 2024 · Table II 本方法 14 筆
表格設定(擷取紀錄原文):Incremental ESDF on the Desktop platform; error is the median absolute voxel-wise difference to a voxelized ESDF ground truth computed from reconstructed voxel centres to the dataset ground-truth surface; the table cites Redwood as [3] although the caption cites [32] (Millane et al., 2024, Table II)
Median ESDF error (m),Redwood · apartment
只並列這張表在相同設定下報告的方法;以「本方法:」開頭者為本頁方法。失敗、未執行與未報告以標記呈現,不是 0。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Millane et al., 2024 在此表設定下報告的數值(author-reported results),只能在同一個比較組內對照,不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Millane et al., 2024, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| nvblox原文提出 | 0.04 m | (Millane et al., 2024, Table II) |
| voxblox本方法 | 0.06 m | (Millane et al., 2024, Table II) |
| Fiesta | 0.05 m | (Millane et al., 2024, Table II) |
Millane et al., 2018 · Table II 本方法 10 筆
表格設定(擷取紀錄原文):CARLA simulated drives l0 and l1 through two synthetic cities; reference geometry is a voxblox reconstruction with ground-truth poses and 0.25 m voxels; evaluated systems use 0.5 m voxels; tracking and integration at 10 Hz; 'Voxblox (ORB-SLAM Poses)' has no dense-map correction after loop closure (Millane et al., 2018, Table II)
RMSE (m),CARLA (simulated) · l0
只並列這張表在相同設定下報告的方法;以「本方法:」開頭者為本頁方法。失敗、未執行與未報告以標記呈現,不是 0。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Millane et al., 2018 在此表設定下報告的數值(author-reported results),只能在同一個比較組內對照,不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Millane et al., 2018, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Voxblox (GT Poses)本方法 | 0.52 m | (Millane et al., 2018, Table II) |
| Voxblox (ORB-SLAM Poses)本方法 | 2.12 m | (Millane et al., 2018, Table II) |
| Ours (subvolume fusion OFF, ablation) | 0.59 m | (Millane et al., 2018, Table II) |
| Ours (subvolume fusion ON)原文提出 | 0.66 m | (Millane et al., 2018, Table II) |
Vizzo et al., 2022 · Table 6 本方法 7 筆
表格設定(擷取紀錄原文):Point-to-point distance (m) between densely sampled maps (KITTI 100,000,000 points, Cow and Lady 1,000,000 points) and the reference cloud; KITTI reference = all KITTI 07 scans aggregated without downsampling, dynamic objects removed with SemanticKITTI labels; Cow and Lady reference from a high-resolution scanner supplied with the dataset; Octomap evaluated through its own point-cloud export (Vizzo et al., 2022, Table 6)
point-to-point distance to reference map, mean,KITTI Odometry · 07 (with space carving)
只並列這張表在相同設定下報告的方法;以「本方法:」開頭者為本頁方法。失敗、未執行與未報告以標記呈現,不是 0。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Vizzo et al., 2022 在此表設定下報告的數值(author-reported results),只能在同一個比較組內對照,不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Vizzo et al., 2022, Table 6)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Octomap (with space carving) | 0.033 m | (Vizzo et al., 2022, Table 6) |
| Voxblox (with space carving)本方法 | 0.497 m | (Vizzo et al., 2022, Table 6) |
| VDBFusion (with space carving)原文提出 | 0.023 m | (Vizzo et al., 2022, Table 6) |
其他比較組
列出其餘 10 個比較組
- Oleynikova et al., 2017 · Text Sec. VI-B1
- Millane et al., 2018 · Table I
- Zhong et al., 2023 · Table II
- Zhong et al., 2023 · Table III
- Vizzo et al., 2022 · Table 2
- Affan et al., 2026 · Table 1
- Oleynikova et al., 2017 · Text Fig. 6 caption
- Vizzo et al., 2022 · Table 4
- Vizzo et al., 2022 · Table 5
- Millane et al., 2018 · Text Sec.V-C
來源
Oleynikova et al., 2017
(2017)Voxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MAV planning2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1366-1373
DOI 10.1109/iros.2017.8202315arXiv 1611.03631程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本:arXiv:1611.03631 (v1 2016-11-11, v2) https://arxiv.org/abs/1611.03631
程式碼:https://github.com/ethz-asl/voxblox(授權:BSD-3-Clause (repository LICENSE))。有公開程式碼不等於已被重現,也不代表目前版本與論文版本相同。